Forecasting spatiotemporal variations in shelter demand leveraging human mobility data: A data-centric multivariate framework

The increasing frequency and severity of natural disasters have challenged the sustainability and resilience of the communities to a greater extent. In the face of such disasters, emergency shelters play a crucial role in providing temporary spaces for residents. An accurate prediction of shelter demand serves as a foundation for emergency planning. However, little attempt has been made in the literature to forecast time-varying demand for shelters during disasters, mainly due to the challenges in data unavailability and demand complexity. The traditional time series models typically fall short of capturing the complex spatiotemporal dependencies in large-scale data. To overcome these shortcomings, we aim to develop a data-centric framework for forecasting spatiotemporal shelter demand leveraging large-scale mobile sensing data. This framework integrates human mobility networks and the dynamic mode decomposition technique to capture the spatiotemporal movements of people in access to shelters. We demonstrate the applicability of the proposed framework by analyzing and predicting the shelter demand of the residents in Harris County (Texas) under Winter Storm Uri, 2021. Our results show the presence of disparities between low-income and high-income neighborhoods in access to shelters. Additionally, the dynamic mode decomposition approach exhibits better predictive performances than the traditional statistical vector autoregression model. Our framework could support informed decision-making in equitable emergency shelter planning by providing more accurate demand forecasting using human mobility data.